Optimizing training set size and marker density for genomic prediction of phytic acid and trypsin inhibitors in soybean (Glycine max.)

Abstract

Genomic prediction accuracy is highly influenced by training set size (TSS) and marker density (MD). However, determining the optimal TSS and MD remains trait and model specific. This study assessed the optimum TSS, MD, and statistical models for predicting phytic acid and total trypsin inhibitors (TTI) using a mid-density SNP panel from 308 soybean genotypes. Parametric models, Genomic Best Linear Unbiased Prediction (G-BLUP), Ridge Regression Best Linear Unbiased Prediction (RR-BLUP), Bayes A, B, C, Least Absolute Shrinkage and Selection Operator (LASSO), and Elastic Net, and non-parametric (Random Forest) models were evaluated across training set proportions (0.5–0.9) and marker densities (0.05–1.0). Random Forest consistently outperformed the parametric models in terms of predictive accuracy. The highest prediction for phytic acid was achieved at an 80:20 training-to-test ratio (R2 = 0.95; RPD = 2.81), whereas TTI prediction peaked at the same ratio (R2 = 0.97; RPD = 2.54). Across traits, TSS between 0.5 and 0.9 exhibited minimal differences, and MD above 50% provided negligible gains in model performance. Optimal predictions were TSS of 0.7–0.8, MD of 1.0 for phytic acid, and TSS of 0.5 MD of 0.8 TTI. These findings emphasize the importance of assessing TSS, MD, and prediction models to specific traits for efficient integration into breeding programs.